面向交互式问答的省略恢复技术研究

交互式问答系统能够通过对话的方式帮助人们获取所需的信息，在教育、医疗、政务及客服等在线咨询服务上得到广泛的应用。然而，由于对话习惯的原因，人们经常使用省略句进行提问及交互，因此正确理解并恢复用户对话中的省略句对于交互式问答系统来说至关重要。在本课题中，我们重点研究交互式问答系统中用户省略问句的恢复技术，针对省略恢复串行模型的错误累积、对话文本的语境信息稀疏、省略恢复的不完备性这三个问题，分别提出省略识别及恢复联合模型、对话主题背景知识图谱以及实体及事件省略恢复统一框架三种方法，实现提高交互式问答系统中的省略恢复性能，进而更加精准地为人们提供在线知识及信息获取服务的目标。在提升人工智能的理解和认知能力方面有着重大的科学意义。

Interactive Question Answering (iQA) system can help people to obtain their required information in a conversational manner. It is widely applied to the online services of education, medical care, government affair and customer service consultation, etc. However, due to the conversation habit of human, they usually use the elliptical sentences to ask and interact with the iQA systems. Therefore, it is crucial to the iQA systems to correctly understand and recover the elliptic sentences in the user dialogue. In this study, we focus on the ellipsis recovery techniques in the iQA system. To address the three key problems of ellipsis recovery ,namely error accumulation in the pipeline model, context information sparseness on the dialogue text and the incompleteness of the ellipsis recovery, we propose the joint model for ellipsis identification and ellipsis component recovery, the background knowledge graph of dialogue topic and a unified framework for entity and event ellipsis recovery to improve the performance of ellipsis recovery in the iQA system. Furthermore, it can accurately render the online knowledge and information acquisition services for human. There is scientific significance in improving the ability of perception and cognition of artificial intelligence.